CN102052998A - Rotor rub-impact acoustic emission signal recognition method - Google Patents

Abstract

The invention discloses a rotor rub-impact acoustic emission signal recognition method based on Hurst index and approximate entropy. Based on the characteristic that the Hurst index and the approximate entropy can represent the physical meaning of the acoustic emission signals of different rub-impact strengths, the Hurst index is combined with the approximate entropy to obtain a combined characteristic parameter for recognizing the rub-impact acoustic emission signal, the method for calculating the rub-impact acoustic emission signal is improved, and a rub-impact acoustic emission signal recognition system based on BP (back-propagation) neural network is established. The method can accurately recognize the acoustic emission signals of different rub-impact strengths.

Description

The recognition methods of rotor rubbing acoustic emission signal

Technical field

The present invention relates to a kind of recognition methods of acoustic emission signal, particularly relate to a kind of rotor rubbing acoustic emission signal recognition methods based on Hurst index and approximate entropy.

Background technology

Between rotating machinery sound part owing to uneven in the operational process, misalign, factor such as thermal flexure takes place bumps that to rub be more common typical fault.Be to improve the efficient of rotating machinery, the tolerance clearance in the rotating machinery design between the sound part reduces day by day, and this makes when the operational efficiency of rotating machinery is improved, between the sound part to bump the problem of rubbing more outstanding.(Acoustic Emission AE) provides a new approach with its unique advantage for bumping to rub to detect with identification in acoustic emission.Bump between the sound part in the rotor-support-foundation system and can cause when rubbing and bump the place's of rubbing sound part elastic strain and produce acoustic emission, the abundant information of rubbing of bumping has been contained in this acoustic emission, have that characteristic frequency is obvious, spectral range is wide, pick-up unit is simple, reliability is high, volume is little, anti-interference is good, be applicable to characteristics such as on-the-spot detections and bump the malfunction monitoring diagnosis of rubbing, make the sound part in the suitable unit running process of acoustic emission bump the identification of the fault of rubbing with acoustic-emission.But, acoustic emission disturbs because being subjected to the various very noisies of rotating machinery generation in service easily, and because the frequency dispersion effect is easy to generate the distortion distortion, it is difficult more that feasible identification to interested bump-scrape acoustic emission signal just becomes in communication process for bump-scrape acoustic emission signal.At present deep not enough to the research of rotor rub-impact acoustic emission recognition method, therefore also fail to bring into play due effect based on bumping of the acoustic emission fault diagnosis of rubbing.Improve and improvement acoustic emission signal analysis ability, research is the acoustic emission source characteristic recognition method more effectively, is to promote the acoustic emission key in application.

Summary of the invention

The present invention mainly can characterize different characteristics of bumping acoustic emission signal physical significance under the intensity of rubbing from Hurst index and approximate entropy, has proposed a kind of recognition methods of bump-scrape acoustic emission signal.

The present invention adopts following technical scheme:

The present invention is based on the rotor rubbing acoustic emission signal recognition methods of Hurst index and approximate entropy, may further comprise the steps:

(1) employing is bumped the acoustic emission experiment device that rubs and is obtained many group bump-scrape acoustic emission signals;

(2), and extract its approximate entropy according to the Hurst index of improved Hurst index extracting method extraction step 1 described bump-scrape acoustic emission signal.And Hurst index and approximate entropy formed the composite character parameter of discerning bump-scrape acoustic emission signal jointly;

(3) the Hurst index and the approximate entropy of the described bump-scrape acoustic emission signal of step 2 are trained as the input vector of BP neural network, obtain training the BP neural network that finishes;

(4) the Hurst index and the approximate entropy of 2 described extractions bump-scrape acoustic emission signal to be identified set by step are entered in the BP neural network that the described training of step 3 finishes, according to the state of rubbing that bumps of the output identification acoustic emission signal of neural network.The corresponding nothing of setting BP neural network output (1,0,0) is bumped and is rubbed, (0,1,0) correspondence is slightly bumped and rubbed, (0,0,1) correspondence is bumped by force and rubbed.

Advantage of the present invention and effect are:

1. improve the computation process of extracting bump-scrape acoustic emission signal Hurst index, made the Hurst exponential quantity more accurate.

2. choose Hurst index and approximate entropy and form the composite character parameter of discerning bump-scrape acoustic emission signal jointly, and combine, thereby improved the discrimination of system effectively with neural network with very strong mode identificating ability.

Other advantages of the present invention and effect will continue to describe below.

Description of drawings

Fig. 1---based on the bump-scrape acoustic emission signal recognition methods overview flow chart of Hurst index and approximate entropy;

Fig. 1 (a)---training stage synoptic diagram;

Fig. 1 (b)---cognitive phase synoptic diagram;

Fig. 2-1---bump the acoustic emission test unit that rubs; Fig. 2-2---bump the structural representation of the device that rubs;

The 1-motor; The 2-gearbox; The 3-shaft coupling; The 4-bearing; 5-clutch shaft bearing seat; 6-bumps the device that rubs; The 7-base; The 8-rotating disk; The 9-axle; 10-second bearing seat; The 21-stator; The 22-support; The 23-bolt.

Fig. 3 (a)～(c)---difference is bumped the acoustic emission waveform figure under the state of rubbing,

Fig. 3 (a) nothing is bumped acoustic emission waveform figure when rubbing, and Fig. 3 (b) slightly bumps acoustic emission waveform figure when rubbing, and Fig. 3 (c) bumps acoustic emission waveform figure when rubbing by force;

Fig. 4 (a)～(d)---certain organizes no bump-scrape acoustic emission signal Hurst exponential fitting figure,

First section fitting a straight line of Fig. 4 (a), second section fitting a straight line of Fig. 4 (b), the 3rd section fitting a straight line of Fig. 4 (c), the 4th section fitting a straight line of Fig. 4 (d);

Fig. 5 (a)～(d)---certain organizes slight bump-scrape acoustic emission signal Hurst exponential fitting figure,

First section fitting a straight line of Fig. 5 (a), second section fitting a straight line of Fig. 5 (b), the 3rd section fitting a straight line of Fig. 5 (c), the 4th section fitting a straight line of Fig. 5 (d).

Fig. 6 (a)～(d)---certain organizes strong bump-scrape acoustic emission signal Hurst exponential fitting figure,

First section fitting a straight line of Fig. 6 (a), second section fitting a straight line of Fig. 6 (b), the 3rd section fitting a straight line of Fig. 6 (c), the 4th section fitting a straight line of Fig. 6 (d).

Embodiment

Be elaborated below in conjunction with the technical scheme of accompanying drawing to invention:

Fig. 1 is based on the bump-scrape acoustic emission signal recognition methods overview flow chart of Hurst index and approximate entropy.This method may further comprise the steps:

(1) employing is bumped the acoustic emission experiment device that rubs and is obtained many group bump-scrape acoustic emission signals;

(2), and extract its approximate entropy according to the Hurst index of improved Hurst index extracting method extraction step 1 described bump-scrape acoustic emission signal.And Hurst index and approximate entropy formed the composite character parameter of discerning bump-scrape acoustic emission signal jointly;

(3) the Hurst index and the approximate entropy of the described bump-scrape acoustic emission signal of step 2 are trained as the input vector of BP neural network, obtain training the BP neural network that finishes; (shown in Fig. 1 (a))

(4) the Hurst index and the approximate entropy of 2 described extractions bump-scrape acoustic emission signal to be identified set by step are entered in the BP neural network that the described training of step 3 finishes, according to the state of rubbing that bumps of the output identification acoustic emission signal of neural network.The corresponding nothing of setting BP neural network output (1,0,0) is bumped and is rubbed, (0,1,0) correspondence is slightly bumped and rubbed, (0,0,1) correspondence is bumped by force and rubbed.(shown in Fig. 1 (b))

Below in conjunction with drawings and Examples, technical solutions according to the invention are further elaborated.

1, the rotor rubbing acoustic emission signal obtains

Bumping of this example rubs the acoustic emission experiment device shown in Fig. 2-1 and 2-2.Be installed in by one and movably bump the device 6 that rubs on the rotor platform base 7 and simulate to realize bumping between sound and rub.Motor 1 rotates by booster engine 2 and shaft coupling 3 driving shafts 9; 9 on axle bumps the device 6 that rubs and is installed between first and second bearing seats 5 and 10 on first and second bearing seats 5 and 10, and axle 9 is through and bumps device 6 stator center of rubbing.Bump several 4 telescopic bolts 23 are installed on the device 6 that rubs, bolt 23 radially faces toward axle 9 centers along rotating shaft, produces to bump by the position of regulating bolt 23 and then adjusting stator 21 disalignments and rubs.

This experimental provision is selected the UT-1000 sensor for use, frequency range 60～1000kHz; Pregain 40dB; The acoustic emission capture card is 18 A/D resolution.Sensor is installed on the clutch shaft bearing seat 5, and in rotating shaft, sampling rate is made as 1MHz to bump Mo Yuan (promptly bumping the device 6 that rubs), and the speed setting of this experimental provision rotor is at 1500r/min.In reality is differentiated, acoustic emission signal can be divided into three classifications: do not have to bump and rub, slightly bump and rub, bump by force and rub.Fig. 3 (a) nothing is bumped acoustic emission waveform figure when rubbing, and Fig. 3 (b) slightly bumps acoustic emission waveform figure when rubbing, and Fig. 3 (c) bumps acoustic emission waveform figure when rubbing by force.

2, the extraction of characteristic parameter

(1) extraction of Hurst index

For the acoustic emission signal time series

The leaching process of its Hurst index is as follows:

A) setting scale length is n, and whole time series is divided into M the subsequence that length is n, and t the sample elements of m subsequence is designated as x _{T, m}, t=1 wherein, 2 ..., n; M=1,2 ..., M.

B) average of m subsequence is:

Then the standard deviation of m subsequence is:

$S_{n,m}=\sqrt{\frac{1}{n}\underset{t=1}{\overset{n}{\mathrm{\Σ}}}{(x_{t,m}-\stackrel{\‾}{x_{n,m}})}^2}---\left(2\right)$

C) extreme difference of m subsequence is:

$R_{n,m}=\underset{1\&le;k\&le;n}{\max }\underset{t=1}{\overset{k}{\mathrm{\&Sigma;}}}(x_{t,m}-\stackrel{\&OverBar;}{x_{n,m}})-\underset{1\&le;k\&le;n}{\mathrm{<figure-callout\; id="1"\; label="min"\; filenames="HDA0000031370580000023.png,HDA0000031370580000031.png"\; state="\{\{state\}\}">min</figure-callout>}}\underset{t=1}{\overset{k}{\mathrm{\&Sigma;}}}(x_{t,m}-\stackrel{\&OverBar;}{x_{n,m}})---\left(3\right)$

Thereby the heavily mark extreme difference statistic of m subsequence is:

${(R/S)}_{n,m}=\frac{R_{n,m}}{S_{n,m}}---\left(4\right)$

D) acoustic emission signal time series Heavily mark extreme difference corresponding to scale length n is:

${(R/S)}_n=\frac{1}{M}\underset{m=1}{\overset{M}{\mathrm{\&Sigma;}}}{(R/S)}_{n,m}---\left(5\right)$

Scale length n reselected and calculate heavily mark the extreme difference statistic accordingly, just can obtain a series of scale length and heavily mark the extreme difference sequence.Among the present invention, the exponential manner of scale length employing 2 is chosen, and promptly gets from 2 ⁱNearest integer is as scale length, i=1 wherein, and 1.02,1.04,1.06 ..., k.The value of k is decided on acoustic emission signal seasonal effect in time series length, and general 2 ^kBe no more than acoustic emission signal seasonal effect in time series length.

E) heavily mark extreme difference sequence (R/S) _nFollowing relation is arranged between its corresponding scale length n:

(R/S) _n＝C·n ^H

(6)

In the formula 6, C is a constant, and H is the Hurst index.By getting double-log:

log(R/S)＝logC+Hlogn

(7)

F) adopt least square fitting just can estimate the Hurst exponential quantity to formula 7.Among the present invention, adopt the mode of piecewise fitting that formula 7 is carried out least square fitting, promptly scale length sequence N is divided into 4 sections by sequential scheduling from small to large, be respectively N ¹, N ², N ³, N ⁴, the heavily mark extreme difference sequence of its correspondence is respectively

Simulation study is found by experiment, and is the most accurate to the 3rd section Hurst index that carries out the least square fitting gained, and promptly the Hurst index that formula 8 is carried out the least square fitting gained is the Hurst index of acoustic emission signal.

$\log {(R/S)}_{{\mathrm{<figure-callout\; id="3"\; label="N"\; filenames="HDA0000031370580000021.png,HDA0000031370580000033.png"\; state="\{\{state\}\}">N</figure-callout>}}^3}=\log C^{*}+H\log N^3---\left(8\right)$

Fig. 4 (a)～(d), Fig. 5 (a)～(d) and Fig. 6 (a)～(d) be respectively certain group do not have bump rub, slightly bump rub, strong bump-scrape acoustic emission signal Hurst exponential fitting figure.The slope of choosing the 3rd section fitting a straight line respectively is as the Hurst index that respectively bumps acoustic emission signal under the state of rubbing, and its result is the most accurate.

(2) extraction of approximate entropy

For the acoustic emission signal time series u (i), i=0,1 ..., N}, the leaching process of its approximate entropy is as follows:

A) preestablish the value of pattern dimension m and similar tolerance limit r, rule of thumb, get m=2 usually, r=0.1～0.25SD (u), the wherein standard deviation of SD (u) expression acoustic emission signal time series { u (i) }.

B) sequence { u (i) } is formed m n dimensional vector n X (i) in order, that is:

X(i)＝[u(i)，u(i+1)...u(i+m-1)]，i＝1～N-m+1

(9)

To the distance between each i value compute vectors X (i) and its complement vector X (j):

$d[X\left(i\right),X\left(j\right)]=\underset{k=0~m-1}{\max }|u(i+k)-u(j+k)|---\left(10\right)$

C) according to given threshold value r (r＞0), to each i primary system meter d[X (i), X (j)]＜ratio of the number of r and this number and total vector number N-m+1, note is done

That is:

D) earlier will Take the logarithm, ask its mean value to all i again, note is made Φ ^m(r), that is:

${\mathrm{\&Phi;}}^m\left(r\right)=\frac{1}{N-\mathrm{<figure-callout\; id="1"\; label="m\; +"\; filenames="HDA0000031370580000023.png,HDA0000031370580000031.png"\; state="\{\{state\}\}">m</figure-callout>}+1}\underset{i=1}{\overset{N-\mathrm{<figure-callout\; id="1"\; label="m\; +"\; filenames="HDA0000031370580000023.png,HDA0000031370580000031.png"\; state="\{\{state\}\}">m</figure-callout>}+1}{\mathrm{\&Sigma;}}}\ln C_i^m\left(r\right)---\left(12\right)$

To m+1, the process of repetition formula 9～formula 12 obtains Φ again ^M+1(r).

E) approximate entropy of this sequence is in theory:

$\mathrm{ApEn}(m,r)=\underset{N\&RightArrow;\&infin;}{\lim }[{\mathrm{\&Phi;}}^m\left(r\right)-{\mathrm{\&Phi;}}^{m+1}\left(r\right)]---\left(13\right)$

In engineering reality, acoustic emission signal seasonal effect in time series length N can not be infinitely great, and when N was finite value, the estimated value that obtains approximate entropy was:

ApEn(m，r，N)＝Φ ^m(r)-Φ ^m+1(r)

(14)

Can calculate bump-scrape acoustic emission signal seasonal effect in time series approximate entropy with formula 14.

3, training and identifying

The Hurst index is a kind of dependence of parameter whether time series has to(for) the time of differentiating, and it thinks that the Hurst index is that 0.5 time series is at random; The Hurst index is in time series between 0 to 0.5 phenomenon of reversing and replying can occur, and presents anti-continuation; The time series that the Hurst index is between 0.5 to 1 presents a kind of continuous stability.According to experimental analysis, do not have that to bump what collect under the state of rubbing mainly be the background mechanical noise, and the background mechanical noise is mainly by motor oscillating and testing table pedestal up-down vibration and produce, the phenomenon that counter-rotating is replied appears in acoustic emission signal, present anti-continuation, so between the Hurst exponential distribution in 0 to 0.5; When taking place slightly to bump when rubbing, can produce several on the time domain and have periodic sudden acoustic emission waveform, this moment, signal trended towards continuation, and it is big that the Hurst index does not have the Hurst exponential quantity of bumping under the state of rubbing; After bumping the mocha weight, regular and lasting bumping rubs and makes acoustic emission signal integral body present a kind of stable trend, and the Hurst index becomes bigger, and is between 0.5 to 1.So the Hurst exponential quantity of acoustic emission signal is in the different scopes, shows that bumping rubs and be in different state of strength.

Approximate entropy mainly is to measure the probability size that produces new model the sequence from the angle of weighing the time series complicacy, and the probability that produces new model is big more, and the complicacy of sequence is big more, and corresponding approximate entropy is also big more.Do not have and to bump that signal mainly is the background mechanical noise under the state of rubbing, signal amplitude is little, pattern is single, and approximate entropy is less; Take place slightly to bump when rubbing, the bumping to rub of excitation makes the background mechanical noise produce new model on original basis, thereby slightly bumped the acoustic emission signal rubbing under, and approximate entropy becomes greatly; Bump by force under the state of rubbing, bump the source of rubbing and produce a large amount of high strength bump-scrape acoustic emission signals, the complicacy of signal significantly improves, and corresponding approximate entropy is also maximum.

Comprehensive above the analysis can be discerned the different states of rubbing that bumps according to represented different physical significances in Hurst index and the residing scope of approximate entropy and this scope.

In the middle of intelligent algorithm, nerual network technique is one and has highly nonlinear ultra-large power system continuous time, its principal character is that the overall situation effect of continuous time nonlinear kinetics, network, large-scale parallel distributed are handled and associative ability, and this makes nerual network technique have widespread use at aspects such as fault diagnosis, pattern-recognition, Intelligent Information Processing.The present invention adopts the identification facility of ripe relatively BP neural network as bump-scrape acoustic emission signal, with Hurst index and approximate entropy with the mode that makes up input vector as neural network.Setting network output (1,0,0) does not rub for having to bump, and rub for slightly bumping (0,1,0), and rub for bumping by force (0,0,1).The hidden layer of choosing neural network comprises 5 neurons, and output layer comprises 3 neurons.Choose respectively many groups do not have bump rub, slightly bump rub, the Hurst index and the approximate entropy of strong bump-scrape acoustic emission signal, the BP neural network is trained as input vector according to the Bayesian normalization method.After training finishes, extract the Hurst index and the approximate entropy of bump-scrape acoustic emission signal to be identified, be entered in the BP neural network that has trained, discern the state of rubbing that bumps of acoustic emission signal according to the output of neural network.

4, experimental analysis

When rotating speed is 1500r/min on experimental provision, gather not have bump rub, slightly bump rub, each 10 groups of strong bump-scrape acoustic emission signals are as the training sample signal.And then gather 3 groups do not have bump rub, 3 groups slightly bump rub, 2 groups of strong bump-scrape acoustic emission signals upset at random, as bump-scrape acoustic emission signal to be identified.Extract the Hurst index and the approximate entropy of training sample signal, and it is trained network as the input vector of BP neural network together.Extract the Hurst index and the approximate entropy of bump-scrape acoustic emission signal to be identified, be entered in the BP neural network that trains and discern, recognition result is as shown in table 1.From the table data as can be seen, with the input vector of approximate entropy as the BP neural network, the neural network that trains can identify all types of rubbing of bumping entirely truely with the Hurst index.

Table 1

Table 2

Table 3

Hurst index or approximate entropy that training sample signal extraction goes out are trained as the input vector of BP neural network separately as the recognition effect of independent characteristic parameter in order to analyze among the present invention Hurst index and approximate entropy bump-scrape acoustic emission signal.After training finished, Hurst index or approximate entropy that bump-scrape acoustic emission signal to be identified is extracted were input in the neural network, and recognition result is shown in table 2, table 3.Table 2 is that BP neural network input vector only is the experimental identification result of Hurst index, and table 3 is that BP neural network input vector only is the experimental identification result of approximate entropy.Data all can not be discerned when the input vector of BP neural network only is Hurst index or approximate entropy entirely truely and bump the type of rubbing as can be seen from table.Consolidated statement 1 can improve discrimination with Hurst index and approximate entropy as the composite character parameter more as can be known effectively.

Claims (5)

1. the rotor rubbing acoustic emission signal recognition methods based on Hurst index and approximate entropy is characterized in that, may further comprise the steps:

Step 1: obtain many group bump-scrape acoustic emission signals;

Step 2: the Hurst index of extraction step 1 described bump-scrape acoustic emission signal, and extract its approximate entropy, and Hurst index and approximate entropy are formed the composite character parameter of discerning bump-scrape acoustic emission signal jointly;

Step 3: the Hurst index and the approximate entropy of the bump-scrape acoustic emission signal that step 2 is obtained are trained as the input vector of BP neural network, obtain training the BP neural network that finishes;

Step 4: the Hurst index and the approximate entropy of the bump-scrape acoustic emission signal that step 2 is obtained are input in the BP neural network that step 3 training finishes, according to the state of rubbing that bumps of the output identification acoustic emission signal of neural network.

2. the rotor rubbing acoustic emission signal recognition methods based on Hurst index and approximate entropy according to claim 1 is characterized in that, the described improved Hurst index extracting method of step 2 is as follows:

A) for the acoustic emission signal time series

Set scale length n, whole time series is divided into M the subsequence that length is n, t the sample elements of m subsequence is designated as x _{T, m}, t=1 wherein, 2 ..., n; M=1,2 ..., M;

B) average of described m the subsequence of calculation procedure a

With standard deviation

Then the extreme difference of described m the subsequence of step a is $R_{n,m}=\underset{1\&le;k\&le;n}{\max }\underset{t=1}{\overset{k}{\mathrm{\&Sigma;}}}(x_{t,m}-\stackrel{\&OverBar;}{x_{n,m}})-\underset{1\&le;k\&le;n}{\min }\underset{t=1}{\overset{k}{\mathrm{\&Sigma;}}}(x_{t,m}-\stackrel{\&OverBar;}{x_{n,m}});$

C), can get the heavily mark extreme difference statistic of described m the subsequence of step a by described standard deviation of step b and extreme difference

The described acoustic emission signal time series of step a then

Heavily mark extreme difference corresponding to scale length n is

Scale length n reselected and calculate heavily mark the extreme difference statistic accordingly, obtain a series of scale length and heavily mark the extreme difference sequence;

D) the described heavy mark extreme difference sequence of step c (R/S) _nFollowing relational expression is arranged: (R/S) between its corresponding scale length n _n=Cn ^H, wherein C is a constant, H is the Hurst index;

E) the described relational expression of steps d is got double-log: log (R/S)=logC+Hlogn, adopt the mode of piecewise fitting that following formula is carried out least square fitting, promptly scale length sequence N is divided into 4 sections by sequential scheduling from small to large, be respectively N ¹, N ², N ³, N ⁴, the heavily mark extreme difference sequence of its correspondence is respectively

To the 3rd segment mark degree length sequences N ³With corresponding heavily mark extreme difference sequence

Carry out least square fitting, promptly right

Carry out least square fitting, the H of gained is the Hurst index of the acoustic emission signal after the improvement.

3. the rotor rubbing acoustic emission signal recognition methods based on Hurst index and approximate entropy according to claim 1 and 2 is characterized in that: the classification of described training sample and bump-scrape acoustic emission signal to be identified comprises: do not have to bump and rub, slightly bump and rub and bump by force and rub.

4. according to claim 1 or 2 or 3 described rotor rubbing acoustic emission signal recognition methodss based on Hurst index and approximate entropy, it is characterized in that, the state of rubbing that bumps of acoustic emission signal is discerned in the described output according to the BP neural network of step 4, set BP neural network output (1,0,0) corresponding do not have bump rub, (0,1,0) correspondence slightly bump rub, (0,0,1) correspondence bumps by force and rubs.

5. the rotor rubbing acoustic emission signal recognition methods based on Hurst index and approximate entropy according to claim 2 is characterized in that: the exponential manner of described scale length n employing 2 is chosen, and gets from 2 ⁱNearest integer is as scale length n, i=1 wherein, and 1.02,1.04,1.06 ..., k, the value of k is decided on seasonal effect in time series length, and 2 ^kBe no more than seasonal effect in time series length.

Images